Potential conflict of interest. Nothing to report.
This study was undertaken to evaluate an image processing method for assessing liver fibrosis in conventional computed tomography (CT) scans in patients with chronic hepatitis C. Two cohorts (designated “estimation,” n = 34; and “validation,” n = 107) of chronic hepatitis C patients were assessed using digitized conventional helical CT. Weighted CT mean fibrosis (Fibro-CT) was calculated as a nonlinear weighted mean F-score for each sample. Fibrosis was defined according to Scheuer on the F0 to F4 scale by 2 pathologists blinded regarding the Fibro-CT data. Fibrosis according to Fibro-CT correlated with histology-determined fibrosis (r = 0.69; P < 0.001) and with increasing F-stage: F0 = 0.23 ± 0.39; F1 = 0.90 ± 0.99; F2 = 1.41 ± 0.94; F3 = 2.79 ± 0.55; F4 = 3.15 ± 0.35 [analysis of variance: P < 0.0001). The receiver operating characteristics curve to diagnose significant fibrosis (≥F2) was 0.83; 95% confidence interval (95%CI), 0.75 to 0.91; and, to diagnose advanced fibrosis (≥F3), was 0.86, 95%CI: 0.80 to 0.93. The correlation between Fibro-CT and fibrosis was higher in patients with homogeneous distribution of fibrosis than in patients with heterogeneous distribution (r = 0.77 versus r = 0.43; P < 0.05). Conclusion: Optical digital analysis of CT images of the liver is effective in determining the stage and distribution of liver fibrosis in chronic hepatitis C. In patients with homogeneous fibrosis distribution, the correlation between Fibro-CT and histology was better than in patients with heterogeneous distribution. Fibro-CT is a simple to use, readily available, and useful method for the diagnosis of fibrosis in patients with chronic hepatitis C. (HEPATOLOGY 2008.)
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
Liver fibrosis stage in chronic hepatitis C defines prognosis and decision-making in treatment.1, 2 The histological evaluation of tissue obtained by ultrasound-guided liver biopsy is accepted as the gold standard in the study of liver fibrosis. However, percutaneous liver biopsy is expensive and is associated with morbidity that, in some cases, can lead to major complications such as bleeding, or death.3, 4 Sampling error and intrapathologist and interpathologist variability5 are the main weaknesses in interpreting fibrosis in liver biopsy specimens.6, 7 Longer (>25 mm) tissue specimens would decrease sampling error as well as interobserver variability, but at the expense of an increased risk of complications.8
There have been developments in imaging techniques over recent years, as well as more specific biochemical analyses for diagnosis and treatment monitoring. Several biochemical markers have been used to estimate the extent of liver fibrosis. These include: (1) routine biochemical markers such as transaminases or platelets9–11; (2) nonroutine surrogate serum markers12; (3) markers of fibrogenesis13, 14; and (4) glycomic and proteomic analysis.15 Image analyses such as those obtained by ultrasonography–elastography,16 Doppler ultrasonography,17 or magnetic resonance18 also have been used successfully. Despite these different approaches, these methods have certain weaknesses in common: (1) external validation is necessary to confirm the diagnosis accuracy; (2) liver biopsy would be avoided in only some patients; (3) the inclusion of fibrosis-related parameters would have little or no effect in the diagnosis accuracy of these tests; (4) these methods can identify patients with advanced fibrosis, but diagnostic accuracy decreases when these methods are applied to medium or low fibrosis stage.19 Elastography has been validated in an external cohort and showed slightly better results than serum markers20 and, when used in combination with Fibrotest, the diagnostic accuracy is improved relative to elastography alone.21
We evaluated liver fibrosis using computed tomography (CT) of conventional images obtained without contrast medium using histogram analysis22 and Fourier transformation.23 This provides a potentially useful tool for image processing by exploiting the computationally effective algorithms of Fast Fourier Transform.24 The purpose of the current study was to evaluate this method of image processing to determine the stage, extent, and distribution of liver fibrosis in conventional CT scans from patients with chronic hepatitis C.
AUROC, area under the receiver operating curve; CT, computed tomography.
Patients and Methods
We studied two cohorts. The first group, termed the “estimation cohort,” comprised 34 patients from the liver unit of the Valme University Hospital (male/female = 20/14; age range, 26-67 years; mean, 46.03 years) with biopsy-proven chronic hepatitis C. All patients had elevated aminotransferases and positive serum hepatitis C virus RNA. Fibrosis distribution was as follows: fibrosis 1 (n = 9), fibrosis 2 (F2: n = 10), fibrosis 3 (F3: n = 9), fibrosis 4 or cirrhosis (F4: n = 6). In all cases, time lapse between biopsy and the analysis of CT images was less than 3 months. All individuals provided written, informed consent for liver biopsy and CT scan. The second group, termed the “validation cohort,” comprised 107 consecutive patients infected with hepatitis C, from 8 Spanish hospitals. All patients were negative for hepatitis B surface antigen and anti–human immunodeficiency virus and accepted to participate in this study. Patients with data supporting co-existence of other liver diseases such as non-alcoholic steatohepatitis, autoimmune hepatitis, or alcohol abuse were excluded. The biochemical and histological characteristics of this validation cohort are summarized in Table 1.
Table 1. Biochemical and Histological Characteristics of the Estimation and the Validation Cohorts
Liver specimens were considered useful for histological diagnosis and fibrosis staging based on length and number of portal tracts. Samples smaller than 10 mm or showing fewer than 10 portal tracts were excluded. Percutaneous liver biopsies were performed under ultrasonographic guidance. Liver specimens were obtained by Tru-cut biopsy (sample length/diameter = 20/1.2 mm) using a biopsy gun and extracting 1 sample per patient. The grading and staging assignments were according to the system of Scheuer et al.25
For Grading: Necro-inflammatory Activity.
aPortal inflammation and interface hepatitis ranging in grade from absent (P0) to severe and widespread interface hepatitis (P4)
bLobular activity ranging from none (L0) to bridging confluent necrosis (L4)
For Staging: Fibrosis.
Staging was performed according to a 5-level scale: F0 = none; F1 = enlarged, fibrotic portal tracts; F2 = periportal or porto-portal septa but intact architecture; F3 = fibrosis with architectural distortion but no obvious cirrhosis; F4 = probable or definite cirrhosis. A further 3-level scale of fibrosis was applied: mild = (F0 + F1), moderate = (F2), and advanced = (F3 + F4).
Hepatocyte steatosis was quantified as the percentage of hepatocytes that contained fat droplets, between 0 (absent) and 100% (all hepatocytes containing fat droplets). To classify the patients, grading of steatosis was assigned as grade 0 if no steatosis, grade 1 when less than 25% of hepatocytes contained fat droplets, grade 2 when between 25% and 50% showed steatosis, and grade 3 when more than 50% of hepatocytes showed fat storage. Patients showing steatosis combined with hepatocyte ballooning, neutrophil infiltrate, and either Mallory hyaline or sinusoidal fibrosis were considered as having steatohepatitis and were excluded from this study.
Liver biopsies were evaluated by the local expert in each of the hospitals collaborating in the study (the fibrosis staging termed “Fhospital”). The liver samples were then referred to a central laboratory where 2 expert pathologists (V.M.C.M. and M.N.S.), blinded with respect to the clinical provenance of the samples, independently evaluated the samples and arrived at a consensus opinion of fibrosis (termed “Fconsensus”) for each patient. Any discrepancies were resolved by discussion between these 2 pathologists. We used the Fconsensus as the reference (gold standard) for subsequent comparisons. The diagnostic accuracy and area under the receiver operating curve (AUROC) of Fhospital was compared with Fibro-CT.
Computed Tomography (CT) Images
Conventional CT scans of the liver as axial slices were obtained without the use of contrast medium in a helical CT apparatus (Toshiba Asteion, at 190-205 mAs, 120 kV, 2 mm beam width, pitch 5.5). Window level/window width (71/71 Hounsfield units) were used to calculate minimum and maximum window values so that each slice could be converted into a numerical matrix of pixels within that window range. Eight radiologists were involved in this study, 1 per collaborating center. CT images obtained under protocol-defined parameters were sent to the referral center for processing. All CT images were processed by 1 researcher (A.M.) under the supervision of the physicist (E.G.G.) responsible for the development of the software. A grid of square samples for analysis was automatically defined on each CT slice. Exclusion of squares from analysis was done by an expert radiologist when any anatomical disturbance was observed. Thus, only squares comprising liver parenchyma were analyzed. The parameter correlating with histological fibrosis was extracted, applying algorithms of Fast Fourier Transform to the different gray shades in each square. To apply this transform, and for numerical filtering, generalized spatial frequencies of Hounsfield units were previously defined in each square.
Processing of Images
Image plates were digitized using the Microtek 9800XL and Adapter TMA 1600 (Microtek Europe B.V., De Hoogvliet—Rotterdam, The Netherlands) with digitization resolution of 300 dpi and depth of 8 bpp, in gray scale. Image files corresponding to each CT slice had a spatial resolution of 0.24 mm/px. Digital image analysis and visualization of results were performed using proprietary software developed by a co-author (E.G.G.), and processed in a conventional computer (processor Intel PIVat 2.67 MHz, 2 Gb RAM) running under the Windows XP operating system.
Per-slice evaluations of CT scans were performed in the following steps: Approximate (interactive) organ segmentation and automated segmentation of liver parenchyma by a combination of thresholding and morphological operations (opening, closing, dilating, and eroding)
Automated Generation of a Grid of Squared Samples of Side l = 20 mm, Overlapped to Segmented Image of Organ.
To improve efficiency and reduce computing time, the grid in the closest approach to the segmented contour was obtained. Because spatial resolution of digitized slices was 0.24 mm/px, and slice thickness (in processed patients) was TS = 4 mm, each digital sample (square) of side LS = 20 mm = 20/0.24 px ≈ 83 px corresponding to a volume of VS = L TS= 1600 mm3 = 1.6 mL liver tissue.
Numerical Analysis of Each Sample Included Generation of Normalized Histogram and Computation of the Power Spectrum in Its Fourier Transformation.
In analyzing this transformation, we defined a “generalized frequency” (horizontal) axis in its “physical” magnitude (f). The equivalent “angular” magnitude (ω) would be defined as ω = 2πf. A band-pass frequency filter (cutoff physical frequency, fo = 0.1934 cycles/Hounsfield unit) was applied, and the area under the curve within a certain frequency range (f > fo) was calculated. This value was defined as the “estimating parameter” (E-value) of the sample analyzed. The calculated F-value of the sample was that given by the defined E - F equivalence (0 < F0,F1 < 0.023 < F2 < 0.040 < F3,F4 < 0.900).
Generation of Color-Coded Map of Calculated F-Values.
Per-patient evaluation was performed by sequentially applying the described process to all slices (n = 6) generated for each patient (Fig. 1).
Definition of Homogeneity in the Distribution of Fibrosis in the Liver
The distribution of fibrosis in the liver was classified in 3 categories: homogeneous, heterogeneous, or intermediate, according to the percentage of fibrosis stage in each processed image. The distribution was considered homogeneous when any stage of fibrosis was represented in more than 75% of squares and no representation was observed of fibrosis more than 1 stage different from the most abundant stage. The distribution was considered heterogeneous when no fibrosis stage was seen in at least 25% of the squares. The rest were assigned as intermediate.
Concordance between numerically calculated F-values (Fibro-CT) and histologically defined fibrosis were compared by Spearman regression coefficient analysis. The kappa coefficient was used to assess concordance between pathologists. Multiple linear regression analysis was used to compare continuous variables. Receiver operating characteristics curves were generated to evaluate the diagnostic accuracy of Fibro-CT as well as the prediction of nonsignificant fibrosis (F0-F1 versus F2-F3-F4) or of advanced fibrosis (F3-F4 versus F2-F1-F0). The statistical method to compare the AUROC curves was based on the method of Hanley and McNeil.26 Sample size was calculated to show a significant difference between histological fibrosis and Fibro-CT. The sample size of the validation cohort was 104 patients with a significance level (alpha) of 0.05, 1 − power (beta) of 0.20, prevalence of significant fibrosis 0.5, and, under the hypothesis of AUROC curve, a difference of less than 0.10 (in other words, AUROC for Fibro-CT, 0.85; and for histological fibrosis, 0.95). Software packages SPSS 15.0 (SPSS, Chicago, IL), EpiDat (EpiDat, Santiago de Compostela, Spain), and NCSS/PASS/GESS 2005 (NCSS, Kaysville, UT) were used for statistical analyses and generating the graphics.
Epidemiological, Biochemical, and Histological Characteristics of the Patients
In the validation group, 13 liver specimens were considered not useful for fibrosis staging because they were less than 10 mm in length or had fewer than 10 portal tracts. Eight CT images had been obtained under conditions that were not defined in the protocol and could not be processed. Thus, 107 CT studies were finally included in the validation cohort and 34 in the estimation cohort.
Fibrosis stage and necro-inflammatory activity were evaluated by the expert pathologist at each participating hospital. Furthermore, all samples were referred for centralized evaluation by 2 pathologists blinded with respect to clinical and Fibro-CT data. When there were discrepancies, a consensus fibrosis was arrived at by discussion. As such, 2 sets of opinions on the fibrosis were defined: Fhospital standard histological fibrosis by the pathologist from the location at which the tissues were generated and the Fconsensus histological fibrosis from the centralized pathology laboratory. There were 107 liver samples classified as optimal with respect to size and portal tract number. Steatosis was absent in 67 cases (62.6%), lower than 25% in 34 cases (31.8%), and higher than 25% in only 6 cases (5.6%). The kappa degree of concordance between the 2 pathologists was 0.72 [95% confidence interval (CI), 0.59–0.84] for mild (F0–F1) fibrosis and 0.75 (95% CI, 0.60 –0.90) for the diagnosis of advanced fibrosis (F3–F4). The diagnostic accuracy of Fhospital as assessed by AUROC was 0.93; 95% CI, 0.85–0.98 in the diagnosis of mild fibrosis (F0–F1) and AUROC, 0.95; 95% CI, 0.92–0.97 for advanced fibrosis.
Prediction of Clinical Significant Fibrosis: Estimation Cohort
A monotonic increase of Fibro-CT values corresponded to the different stages of fibrosis. In the estimation cohort, fibrosis calculated from Fibro-CT correlated with histological fibrosis (Fig. 2; Spearman coefficient r = 0.69; P < 0.001), increasing with F-stage: F0 = 0.41 ± 0.97; F1 = 1.19 ± 0.97; F2 = 1.90 ± 0.92; F3 = 2.61 ± 0.54; F4 = 3.35 ± 0.35 [analysis of variance, P < 0.0001). Clinically significant fibrosis was predicted by Fibro-CT with an AUROC of 0.76 ± 0.09 (95% CI, 0.59–0.93). The AUROC predicting advanced fibrosis was 0.90 ± 0.51 (95% CI, 0.80–0.99).
Prediction of Clinical Significant Fibrosis: Validation Cohort
The AUROC curve for the diagnosis of advanced fibrosis (F3–F4) using Fibro-CT was 0.86; 95% CI, 0.80–0.93 (Fig. 3A). The AUROC for diagnosis of clinically significant fibrosis (F2–F4) using Fibro-CT was 0.83; 95% CI, 0.74–0.91) (Fig. 3B). The presence of steatosis did not influence the ability of Fibro-CT to predict liver fibrosis. However, approximately 95% cases had mild or no steatosis.
Distribution of Fibrosis Homogeneity
Color-coded maps of calculated F values of all 6 slices analyzed enabled the homogeneity of distribution of fibrosis in the liver to be assessed. The fibrosis distribution was arbitrarily classified into 3 categories of homogeneity: 42 patients (39.2%) showed homogeneous distribution, 32 patients (30%) showed heterogeneous distribution, and 33 patients (30.8%) showed an intermediate homogeneity in the distribution of the fibrosis (see Patients and Methods section for category assignment).
The fibrosis distribution influenced the degree of correlation between Fibro-CT and histological fibrosis. In homogeneous liver the Spearman coefficient was r = 0.77 (P < 0.001), in the intermediate was r = 0.49 (P = 0.003), and in patients with heterogeneous distribution the coefficient was r = 0.43 (P = 0.013). The AUROC for the diagnosis of significant fibrosis was 0.89 (95% CI, 0.79–0.99) in patients with homogeneous fibrosis distribution, 0.74 (95% CI, 0.56–0.92) for intermediate distribution, and 0.61 (0.42–0.79) for heterogeneous distribution. Lastly, the AUROC for the diagnosis of advanced fibrosis was 0.98 (95% CI, 0.97–0.99) in homogeneous distribution, 0.74 (95% CI, 0.56–0.92) in intermediate distribution, and 0.61 (95% CI, 0.41–0.80) in heterogeneous distribution.
Comparison Between Standard Histological Fibrosis and Fibro-CT
The diagnostic accuracy for fibrosis was higher (albeit not statistically significantly) using Fhospital than Fibro-CT [AUROC (Fhospital) = 0.91; 95% CI, 0.86–0.96 versus AUROC (Fibro-CT) = 0.83; 95% CI, 0.75–0.91; P = NS]. In the group of patients with homogeneous fibrosis distribution, the diagnostic accuracy was very similar (Fig. 4): AUROC (Fhospital) = 0.86; 95% CI, 0.75–0.97 versus AUROC (Fibro-CT) = 0.89; 95% CI, 0.79–0.99; P = NS.
The main results in the current study support the usefulness of the computerized optical analysis of conventional no-contrast-medium CT scans in the evaluation of liver fibrosis. This method detects the distribution and stage of fibrosis in a full-liver analysis, and the results showed that the correlation between histological fibrosis and Fibro-CT depends on the homogeneity of the fibrosis distribution. Several confounding factors are known to have influenced studies that had addressed the usefulness of noninvasive methods in the diagnosis of liver fibrosis. These factors include: (1) histological analyses of liver biopsy tissue have several limitations such as sampling error and interobserver variability. The length of the specimen obtained is crucial for an accurate diagnosis; shorter samples are associated with fibrosis underestimation. Fibro-CT overcomes these limitations because the method enables full-liver distribution of fibrosis to be assessed. (2) As demonstrated by Regev et at,27 biopsy material obtained from 2 different liver lobes using abdominal laparoscopic exploration showed stage differences in chronic hepatitis C in nearly one third of patients, that is, the distribution of liver fibrosis appears not to be homogeneous, as has been shown by Fibro-CT analysis in the current study. The heterogeneous distribution of fibrosis in the liver could be a significant factor in explaining the inaccuracies associated with noninvasive methods in the prediction of liver fibrosis. Indeed, noninvasive image-based methods such as elastography28 as well as biochemical-based methods29 have not been able to replace completely the requirement for confirmatory percutaneous liver biopsy. (3) A bias frequently found in studies of noninvasive methods for predicting liver fibrosis in chronic hepatitis C has been to consider fibrosis as a quantitative variable derived from qualitative data30 as well as the use of logistic regression as a statistical method to construct equations and scores to predict fibrosis. Recently, neural network analysis has been found to be superior to logistic analysis for predicting liver fibrosis. Fibro-CT was calculated as the mean weighted fibrosis, following the definition of a constant for each fibrosis level, to avoid the difference in fibrosis accumulation being influenced by fibrosis stage. Indeed, cirrhotic liver showed several-fold higher collagen deposition than fibrosis stage 3, whereas such differences were not observed between stages 0 and 1.
Segmentation errors are a key point in defining the limitation of the proposed method. Because the method is based on an analysis of image histograms to quantify subtle differences not visible to the naked eye, the inadvertent incorporation of vessels in the analyzed sample is a confounding factor. In grid squares containing vessels, the extent of fibrosis would be underestimated, whereas other nonparenchyma elements present in the liver would cause the calculated-F to be overestimated. Thus, an experienced radiologist would select the valid images and squares while excluding areas containing vessels, biliary tract, or focal lesions. If liver parenchyma areas are included in selected window intervals (range), then fixing window ranges (level and width) within limits used in visual analysis of CT images would be effective in quantifying the extent of the fibrosis. Massive steatosis could also influence the performance of this method. In the current study, this aspect had not been addressed because of the low number of patients showing hepatocyte steatosis in more than a quarter of hepatocytes. Also, steatohepatitis was an exclusion criterion. Further studies are warranted to assess the relevance of both these aspects. Fibro-CT showed similar diagnosis accuracy as the AUROC curve in the fibrosis evaluation from the individual collaborating hospitals. This indicates the usefulness and reproducibility of this technique. Furthermore, the method enables us to assess the distribution of fibrosis throughout the liver. Transient elastography can explore 1/500th of the liver, whereas liver biopsy assesses only 1/50,000th. Also, the technique has been found useful in predicting fibrosis in several external validation studies that have been conducted to date, mainly in France. The reproducibility of the method is significantly reduced in overweight patients with steatosis, and in cases with a lower extent of hepatic fibrosis.31 The ability to diagnose cirrhosis is similar to that reported for algorithms that include several biochemical-based methods (APRI, Forns, Fibrotest).29, 32 In a recent communication, liver fibrosis had been assessed using magnetic resonance elastography.33 However, there remains a lack of consensus regarding cutoff points to distinguish different stages of fibrosis. Although these methods require further validation and standardization, it would seem feasible that the optimum features for fibrosis assessment would include whole-liver analysis together with elastometry. Several noninvasive methods could be used in succession so that the data derived could be used in combination to improve fibrosis prediction and to avoid the necessity for liver biopsy. Further studies are warranted to assess the correlation between Fibro-CT, transient elastography, and serum-based biomarkers of fibrosis.
In conclusion Fibro-CT emerges as a powerful tool for the diagnosis of liver fibrosis stage in patients with chronic hepatitis C. The accuracy of this method is similar to that achieved by histological evaluation of fibrosis. The advantage is that it does not requires high budgetary investment in expensive apparatus because CT images can be sent worldwide by e-mail in DICOM format files. User-friendly software is already available to analyze the overall liver parenchyma and to define fibrosis stage as well as distribution.
The authors thank Dr. Peter Turner for editorial assistance.